Modify retraining module to support retraining outside of the data validation tool.

[SimonGoring]

Retraining should be its own module using a set format, possibly derived from the parquet files and some other hybrid. This would allow us to update the training using newly added records to Neotoma (for example).

I need to dig through all the files and the notebooks to see how it's all done exactly, but this would be very helpful, because a first pass of some broader sets of articles seems to show a number of false positives.